Supplementary Information for: Separating Fractal and Oscillatory Components in the Power Spectrum of Neurophysiological Signal Haiguang Wen2, and Zhongming Liu*1,2 1 Weldon School of Biomedical Engineering 2 School of Electrical and Computer Engineering Purdue University, West Lafayette, IN, USA *Correspondence Zhongming Liu, PhD Assistant Professor of Biomedical Engineering Assistant Professor of Electrical and Computer Engineering College of Engineering, Purdue University 206 S. Martin Jischke Dr. West Lafayette, IN 47907, USA Phone: +1 765 496 1872 Fax: +1 765 496 1459 Email: zmliu@purdue.edu Coarse Graining Spectral Analysis (CGSA) Yamamoto et al. proposed two versions of CGSA (Yamamoto and Hughson, 1991, 1993). The 1993 version attempted to address an important limitation of the 1991 version when dealing with oscillations at multiple harmonic frequencies. A simplified implementation of the 1991 version was described in (He et al., 2010). In addition to the above-mentioned papers, we elaborate the basis of CGSA and emphasize its limitations as below. CGSA is based on the same model as IRASA. The model is re-stated as below. π¦(π‘) = π(π‘) + π₯(π‘) (S1) where π(π‘) stands for the fractal time-series component, π₯(π‘) , stands for the oscillatory component, π¦(π‘) stands the neural signal that sums up these two components in the absence of noise. CGSA requires resampling the measured time series signal, π¦(π‘), by a factor of h, resulting in a new time series, π¦β (π‘), sampled at 1/h times the original sampling rate. To estimate the PSD of the underlying fractal component (i.e. πΉ 2 (π)), it was proposed to calculate the cross spectrum of π¦(π‘) andπ¦β (π‘), denoted as ππ¦π¦β (π) (Yamamoto and Hughson, 1991, 1993). Considering Eq. S1, we can express ππ¦π¦β (π) as Eq. S2, ππ¦π¦β (π) = [πΉ(π)π ππΌ(π) + π(π)π ππ½(π) ][πΉβ (π)π −ππΌβ (π) + πβ (π)π −ππ½β (π) ] = πΉ(π)πΉβ (π)π π(πΌ(π)−πΌβ (π)) (1 + Ψβ (π)π −ππ(π) )(1 + Ψβ (π)π ππβ (π) ) (S2) where Ψ(π) = π(π)⁄πΉ(π), Ψβ (π) = πβ (π)⁄πΉβ (π), π(π) = πΌ(π) − π½(π), θβ (π) = πΌβ (π) − π½β (π). Note that Ψ and θ indicate the relationship between the oscillatory and fractal components in terms of their ratio in magnitude and their difference in phase, respectively. If π¦(π‘) only contains the fractal component (i.e. π(π) and πβ (π) both equal to zero for any π), the magnitude of the cross spectrum, βππ¦π¦β (π) β, can be simply expressed as Eq. S3, βππ¦π¦β (π) β = πΉ(π)πΉβ (π) = βπ» πΉ 2 (π) (S3) If π¦(π‘) only contains a simple oscillation with a single frequency π0 (i.e. πΉ and ππβ both equal zero for any π), it is straightforward to show that Eq. S4 holds true. βππ¦π¦β (π) β = 0 (S4) It is thus tempting to estimate the PSD of the fractal component by cancelling h with Eq. S5. πΉ 2 (π) = √βππ¦π¦β (π) β βππ¦π¦1/β (π) β (S5) Eq. S5 is central to CGSA in an attempt to estimate πΉ 2 (π) independent of the resampling factor h. However, it is important to note that Eq. S5 cannot be generalized to more realistic cases in which π¦(π‘) includes both fractal and oscillatory components, because the above cross power spectrum bears very complicated interactions between the fractal and oscillatory components according to Eq. S2. In the following, we will show that such interactions remain difficult to eliminate when the cross-spectral analysis is used in an attempt to extract the fractal component. To start with the perhaps simplest case, we assume that π₯(π‘) is a simple oscillation with a single harmonic frequency π0 . Therefore, π(π) or πβ (π) is non-zero only at π0 or π0 β , respectively. Based on Eq. S2, we can rewrite βππ¦π¦β (π) β as βπ» πΉ 2 (π), π ≠ π0 and π ≠ π0 β π» 2 (π)β1 + Ψ(π)π ππ(π) β, π = π0 ππ¦π¦β (π) = { β πΉ (π) π» 2 (π)β1 ππ β β πΉ + Ψβ (π)π β, π = π0 β (S6) And the estimate of the PSD of the fractal component is expressed as Eq. (S7). πΉ 2 (π), π ≠ π0 , π0 β and π0 /β 2 (π)β1 πΉ + Ψ(π)π −ππ(π) β, π = π0 √βππ¦π¦β (π)β βππ¦π¦1/β (π)β = πΉ 2 (π)√β1 + Ψβ (π)π ππβ (π) β, { π = π0 β (S7) πΉ 2 (π)√β1 + Ψ1/β (π)π ππ1/β(π) β, π = π0 /β Eq. S7 suggests that the estimated power spectrum deviates from the scaled power-law distribution at not only π0 but also π0 β and π0 ⁄β, giving rise to the residual oscillation and the processing artifact in the estimated PSD of the fractal component (also see Fig. 2 in Yamamoto and Hughson 1993). At π = π0 , we can quantify the relative error, denoted as π πΈ(Ψ, π), using Eq. S8 π πΈ(Ψ, π)|π=π0 = β1 + Ψ(π0 )π−ππ(π0) β − 1 (S8) This relative error depends on the gross degree of interaction between the oscillatory and fractal components in terms of their ratio in magnitude and phase difference at π = π0 . It is worth noting that this error will not vanish by averaging the cross power spectra estimated from multiple time segments of the original time series (Yamamoto and Hughson, 1991, 1993). In CGSA, one may assume a random phase relationship between the oscillatory and fractal components. That is, their phase difference π(π) for any individual time segment follows a uniform random distribution in [0, 2π]. For the sake of simplicity, let us further assume that πΉ(π) and π(π) remain stationary across different time segments. Thus the relative error resulting from averaging over N time segments can be expressed as 1 −πππ (π0 ) β−1 π πΈ(Ψ, π)|π=π0 = ∑π π=1β1 + Ψ(π0 )π π (S9) Given the assumed uniformly random phase distribution, it can be shown that the relative error does not necessarily converge to zero even with a sufficient number of time segments for averaging. Instead, it converges to a value that increases with the increasing ratio in magnitude between the oscillatory and fractal components at π = π0 . The relative error at π = π0 β follows the similar trend (see the simulation results in Fig. S1). Moreover, the theoretical limitations of CGSA have more profound effects when the oscillatory component has multiple harmonic frequencies. This reflects a more realistic circumstance, as neural signals are rarely strictly sinusoidal with a single frequency. For example, let us suppose the oscillatory component to have a fundamental frequency at π0 and a second harmonic frequency at 2π0 . If one resamples the neural signal by a factor of h=2, the original oscillatory component and its resampled version overlap in the spectral domain. As a result, their cross power spectrum is non-zero at the harmonic frequencies, resulting in an even larger relative error as expressed in Eq. S10 than is expressed in Eq. S8. π πΈ(Ψ, π)|π=π0 = β1 + Ψ(π0 )π−ππ(π0) β√β1 + Ψβ (π0 )πππβ (π0 ) ββ1 + Ψ1/β (π0 )πππ1/β(π0 ) β − 1 (S10) Importantly, this additional error tends to occur when resampling the signal by an integer factor (or its reciprocal) and becomes more serious when the oscillatory component has more harmonic frequencies. A modified CGSA algorithm was proposed to address this problem by using a phase orthogonalization method (Yamamoto and Hughson, 1993). In this modified algorithm, the phase of the fractal component is further assumed to be random and uniformly distributed within [0, 2π], whereas the phase of the oscillatory component is assumed to evolve in a stationary manner. It is important to note that the phase orthogonalization method (Yamamoto and Hughson, 1993) is not intended to address or eliminate the interference between the fractal and oscillatory components. Instead, it attempts to eliminate the cross spectrum between the oscillatory component and its resample version when the oscillatory component contains multiple harmonics. Mathematically, let us rewrite Eq. S2 as Eq. S11. ππ¦π¦β (π) = πΉ(π)πΉβ (π)π π(πΌ(π)−πΌβ (π)) + π(π)πβ (π)π π(π½(π)−π½β (π)) +πΉ(π)πβ (π)π π(πΌ(π)−π½β (π)) + π(π)πΉβ (π)π π(π½(π)−πΌβ (π)) (S11) Obviously, the phase orthogonalization tends to minimize the second term, but does not necessarily reduce the third or fourth term in Eq. S11. This is in part because the relative phase difference between the (original/resample) fractal and (resampled/original) oscillatory components are random and non-stationary. In short, the complex interactions between fractal and oscillatory components remain problematic despite the use of phase orthogonalization and impede using the cross-spectral analysis to separate the oscillatory and fractal components as shown by our simulation results (Fig. 3). Moreover, the aforementioned assumptions for this phase correction method to be applicable may not be valid in practice and likely confound the result. Importantly, though the irregular-sampling technique can successfully remove the oscillatory component from fractal component in IRASA, the same strategy is not useful in CGSA. This is mainly because in CGSA the relative errors at the oscillatory frequency, π0 , are not relocated with different h values. So for any h value, the cross-spectral power at the oscillatory frequency always deviates from the power-law distribution, which is an inevitable bias that cannot be addressed with the same statistical method as used in IRASA (see Fig. S2 for an illustrative example). For a fair performance comparison, we employed multiple h values for both the proposed IRASA method and the modified CGSA (Yamamoto and Hughson, 1993) followed by taking the median across results obtained with different resampling factors. To quantitatively evaluate and compare IRASA and CGSA, we defined the percentage error in the estimated power-law exponent and the mean square error (MSE) in the estimated fractal power spectrum by using (S12) and (S13), respectively. The power-law exponent was estimated from the slope of the linear function best fitting the extracted PSD of the fractal component. πππππ = |π½ππ π‘ππππ‘ππ −π½π‘βπππππ‘ππππ | π½π‘βπππππ‘ππππ × 100 (S12) πππΈ = avg〈|log πππ·ππ π‘ππππ‘ππ − log πππ·π‘βπππππ‘ππππ |2 〉 (S13) π As shown in Fig. S3, the relative errors in the estimated fractal PSD and power-law exponent were both significantly smaller for IRASA than for CGSA. As the number of oscillations increased, CGSA had a deteriorating performance, whereas IRASA behaved robustly. We also noticed that the performance of CGSA depended heavily on the phase distribution of the fractal component. Letting the fractal phase randomly vary from 0 to ππππ₯ , we found that the errors for CGSA increased almost exponentially as ππππ₯ became smaller (i.e. the fractal phase became less random), whereas the performance of IRASA was nearly independent of the range of the fractal phase distribution. And IRASA is much more robust against the relative amplitude of the oscillatory component to the fractal component. These simulation results suggest that IRASA is superior to CGSA. Fig. S1 Relative error in the estimated power spectrum of the fractal component at frequency π0 for CGSA (A) The relative error converges to non-zero values as the number of averages increases. (B) The converged relative error increases with the increasing ratio in magnitude between the oscillatory component and the fractal component (i.e. Ψ). (C) The relative error also depends on the range of the phase distribution (i.e. [0, ππππ₯ ]) Fig. S2 Extracting the fractal power by using multiple h values in IRASA and CGSA. (A) The simulation signal is composed of a fractal signal (π½ = 1.5, Ψ = 2, ππππ₯ = 2π) and an oscillatory signal at π0 = 20π»π§. (B) The extracted fractal powers corresponding to four different h values (1.2, 1.4, 1.6 and 1.8). (C) Estimate the fractal power by taking the median from the power spectra of multiple h values. Note that in IRASA, the oscillatory power would appear as statistical outliers, the effect of which can be removed by taking the median. However in CGSA, the oscillatory power would always be present at the oscillation frequency despite the use of different resampling factors Fig. S3 Performance evaluation of IRASA and CGSA. (A) Effects of the number of oscillation frequencies (π½ = 1.5, Ψ = 1, ππππ₯ = 2π), (B) the amplitude ratio of the oscillatory to fractal component (π½ = 1.5, N = 10, ππππ₯ = 2π), and (C) the range of the fractal phase distribution (π½ = 1.5, N = 10, Ψ = 1), on the errors in the estimated power-law exponent (π½) (top) and the estimated power spectrum of the fractal component (bottom)